Early seizure detection in childhood focal epilepsy with electroencephalogram feature fusion on deep autoencoder learning and channel correlations

Recognition of epileptic electroencephalogram (EEG) signals is vital to epileptic seizure detection. Current research on seizure detection mostly focused on generalized seizure analysis. Compared with generalized seizures, childhood focal epilepsy generally originates in one hemisphere of the brain, and the seizure onset patterns vary from patient to patient, making it difficult for analysis. Meanwhile, the frequency, amplitude and rhythm of EEG activities in children of different ages are different, making childhood focal epilepsy detection challenging. In this paper, the channel correlation features (CCF) containing the regional scalp EEG cross correlations and auto-correlations are proposed for children focal epilepsy EEG representation. The spectrum features are also extracted to characterize EEGs in different frequency bands. Further, a convolutional autoencoder (CAE)-based deep feature learning and dimensionality reduction model is proposed for discriminative EEG frequency domain feature extraction. An early seizure detection framework for the childhood focal epilepsy based on the fused CAE and CCF EEG features is finally developed, and an ensemble classification model (ECM) is applied to solve the imbalance issue between ictal, interictal, and preictal. The performance is evaluated on the EEG dataset collected by the Children's Hospital, Zhejiang University School of Medicine (CHZU). Experiments show that the proposed algorithm can reach to the highest accuracy of 93.57% for the early seizure detection in childhood focal epilepsy.

Automated summary

The research proposed a method for representing childhood focal epilepsy EEG based on channel correlation features and spectrum features, and developed an early seizure detection framework based on CAE and CCF EEG features, solving the issue of imbalance between seizure types using ECM.